Research  ›  Case library  ›  Case 04: Goodhart’s Law & Campbell’s Law
Case 04 / 31  •  Theme: Optimising

Goodhart’s Law & Campbell’s Law

Once a measure steers behaviour, it changes from an observation tool into part of the system.

The popular slogan “when a measure becomes a target…” is broader than Goodhart’s original warning. Goodhart addressed unstable statistical relationships under policy control; Campbell emphasised corruption pressure and distortion when indicators acquire consequences.

✓
Research statusResearch in preparation
☆
AI relevanceAI may amplify an existing effect
⚙
Core ideaProxy–objective divergence
⌾
Evidence baseSources and limitations summarised; independent review pending.
◇
Case typeLaw / methodological warning
“
In one sentence

Once a measure steers behaviour, it changes from an observation tool into part of the system.

The storyHow the case arises and how it can mislead our thinking.

An organisation selects a measurable indicator because it is associated with an important objective. Once bonuses, rankings or automation are tied to that indicator, actors discover ways to improve the figure without improving the underlying objective to the same extent.

What do we know?Evidence, research, critical insights and what remains unknown.

Education, healthcare and performance management show gaming, tunnel vision and score inflation, but targets can also produce real improvement. These laws are therefore not a prohibition on measurement; they are warnings about feedback and design.

What is the mechanism?Which explanation best fits what we observe.

Optimisation pressure can make previously passive proxies part of the changing system. If actors exploit ways to improve the proxy without proportionate improvement in the objective, the relationship between proxy and objective can weaken. Whether this happens, and to what extent, depends partly on incentives, adaptation and independent checks.

What does AI change?How AI may amplify the mechanism or make it visible.

AI transfer to investigate: AI does not create this mechanism, but increases the search power, speed and scale with which specification gaps and loopholes can be found. Reward hacking and specification gaming are direct technical relatives.

What do we want to test?The specific question we want to investigate further.

Proposed test, not yet conducted: Compare informative KPIs, human target pressure, AI optimisation against the same proxy, and AI optimisation with hidden or rotating independent audits. The main measure is the rate of proxy improvement versus independently measured improvement in the objective.

What do we take away?What this case provisionally means — and where its limits lie.

A metric must be reassessed once it becomes an intervention. Good measurement quality before optimisation pressure does not guarantee the same evidential meaning afterwards.

A proxy is a measurable stand-in for an objective. Optimisation pressure can change the relationship; deterioration is not inevitable. Measure the objective independently of the rewarded proxy.

SourcesPrimary and key scholarly sources.

Goodhart (1984), Monetary Theory and Practice; reprint of Problems of Monetary Management

Campbell (1979), Assessing the impact of planned social change

Goodhart (1984), Problems of Monetary Management: The UK Experience — Chapter-specific DOI; full text unavailable; not used as passage evidence

Campbell (1976), Occasional Paper8, pp.49–54 — 1976 working paper, distinct from the 1979 journal article

The passage on corruption pressure and independent checks comes from Campbell’s 1976 working paper (pp. 49–54), not the 1979 journal version. Bibliographic details and context are available for Goodhart’s chapter; its full text was not accessible. Deterioration is not inevitable.

Give feedback

Diagram at full size